everyone: dont build anything serious on claude fable 5
me: ~5 billion tokens deep at 1am, watching claude code finish the ai agent that quietly automates the hardest part of my own forensics job 💀
its called VERDICT, demo below 👇
#claudecode#fable5#dfir
i built an ai that caught a hacker hopping across 6 computers in the same second. then i made it prove every word. digital forensics, open source, and you can check every claim yourself. code: https://t.co/BM7RkZ7DyP demo (5 min): https://t.co/V7gMTdNJx2
every claim points back at the exact tool output it came from or a verifier deletes it, and the whole run is signed so anyone can verify it offline. open source, apache 2.0: https://t.co/BM7RkZ7DyP
roast it, tell me where youd expect it to be confidently wrong.
everyone: dont build anything serious on claude fable 5
me: ~5 billion tokens deep at 1am, watching claude code finish the ai agent that quietly automates the hardest part of my own forensics job 💀
its called VERDICT, demo below 👇
#claudecode#fable5#dfir
the part im actually proud of: two agents argue with each other. one tries to prove the box is hacked, one tries to prove its clean, and they have to agree before anything counts. kills the confident-wrong answers you get when one model just agrees with itself.
Well, this is awkward. I think I just automated the hardest part of my own job.
A few weeks building an AI agent for digital forensics, and somewhere around 3am it hit me: I replaced myself before I figured out how to actually sell the thing.
It's called VERDICT 🧵
The part I'm proud of: it stays honest. On a case with a known answer key it got 5/5, reproducible offline, and it never claims more certainty than the evidence allows. It still hands the final call to a human. So my job is safe. Probably.
Meet VERDICT: point it at digital evidence memory, EVTX, disk, or packet capture and it returns a signed verdict (is there evil here?). Every finding cites the exact tool call, in a chain of custody you can verify offline. A DFIR agent built on Claude Code. #DFIR#SANS
built an AI agent that does what takes SOC analysts 30 min in about 30 seconds. ES|QL correlation, beaconing detection, lateral movement tracing, process chain analysis all mapped to MITRE ATT&CK and wired through @elastic Agent Builder agent finds it every time.
@elastic_devs
Check out our new work "Score-Guided Diffusion for 3D Human Recovery", a.k.a. ScoreHMR, with @ligongh and Dimitris Metaxas that will appear at #CVPR2024!
Paper: https://t.co/lKMFBh2eow
Project Page: https://t.co/arDU2sMDKh
Code & models: https://t.co/xlO0FjYqKz